Exploring the Structure and Function of Brain Networks IAP 2006, September 20, 2006 Yoonsuck Choe Brain Networks Laboratory Department of Computer Science Texas A&M University choe@tamu.edu, http://research.cs.tamu.edu/bnl 1
Research Goal I f The Brain Network Lab s goal is to understand: Organizational principles underlying brain structure. Computational principles underlying brain function. 2
Motivation Deeper understanding of brain organization and function will enable us to: Build new kinds of computing artifacts, more robust and more autonomous (cf. autonomic computing) Allow new kinds of interface between humans and computing systems. Help treat neurological and mental disorders. 3
Approach Structural: Computational neuroanatomy Knife-Edge Scanning Microscope Functional: Computational neuroscience Autonomous semantic grounding Delay in nervous system to delay compensation and prediction 4
Structure of Brain Networks 5
Knife-Edge Scanning Microscope (a) KESM (b) Cutting/Imaging (c) Illumination Section and image mouse brain (15mm 12mm 9mm). Cut thickness: 500nm; x-y Resolution 250nm 250nm. 44 KHz lines/s 4096 pixels = 180MB/s. 20TB of raw data per brain (can be compressed). US Patent 6,744,572: Bruce McCormick, former director 6
KESM Data: Nissl and Golgi Staining 7
3D Reconstruction Image processing and segmentation algorithms. 3D reconstruction and compression algorithms. 8
On-Going and Future Work Distributed storage architecture for massive data. XML-based representation for storage/retrieval, visualization, and analysis. Estimation of temporal properties: axon length, axon thickness, etc. Inferring neuronal network connectivity and property. Mining for basic circuits (modules). 9
Function of Brain Networks 10
Why Do We Have a Brain? (a) Free-floating (b) Settled Marine tunicate (a small marine animal): Has a brain when free-floating. Digests its own brain once settled. Brain is for motion and prediction. 11
Two Questions Regarding Brain Function Grounding: One s own, not programmed, meaning. Motor act plays an important role. Emergent goals: One s own goals/tasks. Prediction/anticipation is a prerequisite. 12
The Problem of Grounding What does the internal representation mean? How can we, as scientists, know the meaning? How can we, as the brain itself, know the meaning? 13
I Grounding, from Within and Without S S f I f (a) External observer (b) Internal observer How can the brain understand its own internal state? Seems possible to an external observer, with access to both the I and S. Seems impossible as an internal observer, with access to only S, but that can t be right. What s missing here? 14
Potential Answer: Action The brain does not passively perceive. The brain generates action. But how and why can that help? 15
Some Experimental Clues Video to tactile array. Tactile when passive. Vision-like when active. Perception as related to voluntary/intentional action may be the key! Bach y Rita (1972; 1983) 16
Key Idea: Action Related to Invariance in the Internal State Visual Environment Visual Field Perception I Filter Bank f Sensor Array s π Action Vector a Eye movement t=1 t=2 t=3 t=2 t=1 t=3.... time Action (a) Sensorimotor agent (b) Invariance during action Movement in a certain direction (diagonal) causes the sensory array to stay invariant over time. Property of such a movement exactly reflects the property of the input I (i.e., oriented diagonally). 17
Sensory Array Activation Raw Input DoG filtered Input Image Sample 1 I R * = I D I Orientation Filters = Response Vector 2...... θ A series of sensory filters, resulting in orientation 18 G r n
Learn Invariance-Preserving Action A: direction of motion S: sensory state (orientation) 0.5 0 0 0 0.5 0 0.5 0 0 0 0 0 Q(s i,a j ) 0 0 0 0 0 0.5 0 0 0 0 0.5 0 0 0 0.5 0 0 0 0.5 (a) Q-table (b) Ideal (c) Learned Learn action that maintains invariance in sensory state, through reinforcement learning. Property of learned action closely reflect that of the corresponding sensory state! 19
Gaze Behavior (a) Initial (b) Final By simply trying to act while maintaining invariance in internal state, one can infer external world properties without peeking outside. This could form the basis for autonomous grounding. 20
Gaze Behavior in Humans Scanpath (Yarbus 1967). Yarbus (1967) Other related work: Mirror neurons (Rizzolatti et al. 1996). 21
Summary: Autonomous Grounding Action plays a key role in autonomous grounding. Preserving invariance in internal state is a simple yet powerful objective. Asking questions from the brain s point of view can help clarify hidden issues. 22
Function of Brain Networks Emergent goals 23
Emergent Goals and Prediction In order to formulate a goal, we need to see into the future. Thus, prediction becomes the first necessity. Prediction links the past to the future. Main research question: Whence prediction? Some kind of extrapolation process? 24
Delay in the Brain: Need for Extrapolation Thorpe and Fabre-Thorpe (2001) Due to neural conduction delay (couple of 100 ms), we cannot even seem to catch up with the present. May need some extrapolation. 25
Extrapolation in Visual Perception t=1 t=2 t=3 t=4 Perceived Flash-Lag Effect (Nijhawan 1994) suggests that the brain may be performing extrapolation to compensate for delay. 26
With Delay Compensation: FLE s ( t ) s ( t + t ) t t t + t Match t, p ( t ) = s ( t ) t + t p ( ) s ( t + t ) 27
Research Questions How can the nervous system compensate for internal delay? Are there single-neuron-level mechanisms for that? 28
Potential Answer (Markram et al. 1998) Facilitating synapses can generate extrapolative activity. 29
Results from Facil. Syn. Model Periphery Periphery 1 1 Spike 0.5 Spike 0.5 0 0 100 200 300 400 500 600 Time 0 0 100 200 300 400 500 600 Time Delay Input (presynaptic) spikes Input (presynaptic) spikes 1 1 Spike 0.5 Spike 0.5 0 0 100 200 300 400 500 600 Time 0 0 100 200 300 400 500 600 Time Postsynaptic membrane potential 350 Postsynaptic membrane potential 350 300 300 250 250 Voltage (mv) 200 150 Voltage (mv) 200 150 100 100 50 50 0 0 100 200 300 400 500 600 Time 0 0 100 200 300 400 500 600 Time Output (postsynaptic) spikes Output (postsynaptic) spikes 1 1 Spike 0.5 Spike 0.5 0 0 100 200 300 400 500 600 Time (a) Increasing firing rate 30 0 0 100 200 300 400 500 600 Time (b) Decreasing firing rate
Application to Real-Time Control z f x f y θ z x θ x y c x c y z x 2D pole balancing problem: Delay introduced in input (position and pole angle). Recurrent neural network controller with facilitatory dynamics. 31
Experiment Compare task performance under three types of dynamics: Control: Basic ESP implementation. FAN: Facilitatory Activation Network. DAN: Decaying Activation Network. 32
Results: Success Rate Success rate $&%' (&)*,+ -.)/ 0%'! "! " # No delay Delay Experiments Delay in θ z Delay in θ x Different delay conditions were tested. FAN showed best performance under all conditions (t-test, p < 0.005, n = 250). 33
Results: Effect of Increased Delay 1 DAN Control FAN 1 DAN Control FAN 0.8 0.8 Success rate 0.6 0.4 Success rate 0.6 0.4 0.2 0.2 0 0 1 step 2 steps 3 steps 0 40 50 60 70 80 Delay in θ z Blank-out duration (a) Increased delay (b) Increased blank-out duration Performance under increased delay and input blank-out period. In all conditions, FAN performed the best. 34
Blank-Out as External Delay Mehta and Schaal (2002) Input feed cut off for 40 500 ms while balancing a virtual pole. Humans are good at dealing with input blank-out. FAM shows similar robustness. 35
Summary: Predictive Function Delay is inevitable, and needs to be compensated. Extrapolation, through facilitation, is a possible mechanism. Extrapolation predicts the present, based on the past. The same mechanism, if pushed a bit farther, can predict the future. 36
Overall Conclusion Structure and function of brain networks: Structural information can provide critical parameters for functional models (connectivity, timing, etc.). Functional models can provide theoretical framework for analyzing the structural information. 37
Credits KESM Project: Bruce McCormick, John Keyser, Louise Abbott, David Mayerich, and Jaerock Kwon. Funding provided by NIH/NINDS, NSF, Texas ARP, TAMU VPR office, and TAMU CS. Autonomous Grounding: S. Kumar Bhamidipati, Noah Smith, and Huei-Fang Yang. Facilitating Activation: Heejin Lim. 38
References Bach y Rita, P. (1972). Brain Mechanisms in Sensory Substitution. New York: Academic Press. Bach y Rita, P. (1983). Tactile vision substitution: Past and future. International Journal of Neuroscience, 19:29 36. Choe, Y., and Smith, N. H. (2006). Motion-based autonomous grounding: Inferring external world properties from internal sensory states alone. In Gil, Y., and Mooney, R., editors, Proceedings of the 21st National Conference on Artificial Intelligence. 936 941. Markram, H., Wang, Y., and Tsodyks, M. (1998). Differential signaling via the same axon of neocortical pyramidal neurons. Proceedings of the National Academy of Sciences, USA, 95. Mehta, B., and Schaal, S. (2002). Forward models in visuomotor control. Journal of Neurophysiology, 88:942 953. Nijhawan, R. (1994). Motion extrapolation in catching. Nature, 370:256 257. Rizzolatti, G., Fadiga, L., Gallese, V., and Fogassi, L. (1996). Premotor cortex and the recognition of motor neurons. Cognitive Brain Research, 3:131 141. Thorpe, S. J., and Fabre-Thorpe, M. (2001). Seeking categories in the brain. Science, 291:260 263. Yarbus, A. L. (1967). Eye Movements and Vision. New York: Plenum. Translated from Russian by Basil Haigh. Original Russian edition published in Moscow in 1965. 38-1